Seite - 134 - in Intelligent Environments 2019 - Workshop Proceedings of the 15th International Conference on Intelligent Environments

In addition to these statisticalmeasureswewillmeasure the number of network errors. Althoughwe are in regression, two types of errors will be distinguished. Error type 1 (Error1)occurswhentheactual temperature ispositiveandyet theLSTMpredictsaneg- ative temperature. Error type 2 (Error2) occurswhen the actual temperature is negative and the LSTMpredicts a negative temperature. Type 1 error despite being an error is less serious, as itmeans that theLSTMpredicts a temperaturebelowzeroand theactual temperature is positive. The same can be extrapolated to any other threshold tempera- ture.However in error type 2 theLSTMpredicts a positive temperature and the actual temperature is negative, in this case the farmer could lose thewhole crop. In the results wearegoing to showthe%ofeachoneof these typesoferrors, beingdefinedas: • %Error1:Percentageof type1errors considering thenumberof test instances. • %Error2:Percentageof type2errors considering thenumberof test instances. 4.1. Adjusting theLSTM In this experiment the adjustment of the LSTM is shown by means of a 3-fold cross validation. Thus, the 17,212 are randomly divided into 3 subsets ofwhich twoof them are trainedand the third isperformed the test, this is repeateduntil youhave testedwith the 3 subsets. Each subset of the test consists of 5,736measures, eachmeasure being considered as an instance.Table 2 shows themean results of the3-fold cross validation experiment. Dataset RMSE MAE PCC R2 %Error1 %Error2 3-cv 0.77 0.41 0.9922 0.9843 0.854% 0.964% Table2. Meanresultsobtainedafter theexecutionofexperimentof3-foldcrossvalidation As it can be seen in table 2, the results obtained for this experiment are quite sat- isfactory since theLSTMobtains anRMSEvalue less than one degreeCelsius and the goodnessof themodel tomake thepredictions isclose tobeingaperfectfit, thevalueof R2 isalmost1.The%of thedifferent typesoferrors is less than1%, therefore the result is acceptable. Figures1,2and3showgraphicallyhowtheprediction trend ispractically thesame as the actual air temperature. Figure 1 refers to the first fold of the experiment, Figure 2 corresponds to the fold and Figure 3 to fold 3, each fold being the individual test of the experiment of the 3-fold cross validation. For the three figures only the first 200 instances of the test are shown in order to be able to visualize correctly the behavior of the predictions with the real temperature. For these three figures only the first 200 instances of the test are shown inorder to be able to visualize correctly the behavior of thepredictionswith the real temperature. 4.2. Predicting the temperatureof24hours In this experimentwe are going to use 10%of the instances to perform the test.As the predictions are not going to be randommeasurements butwe are going to predict full days, 12dayshavebeen selected topredict their temperatures every10minutes in total 1,728measurements(instances).Table3shows the resultsof this experiment. M.Á.Guillén-Navarroetal. /AnLSTMDeepLearningScheme forPredictionofLowTemperatures134